1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/avx2-rr1-p5.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x8(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x8(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n % sizeof(float) == 0);
25
26 const __m256 vsign_mask = _mm256_load_ps(params->avx2_rr1_p5.sign_mask);
27 const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p5.magic_bias);
28 const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p5.log2e);
29 const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p5.minus_ln2);
30 const __m256 vc5 = _mm256_load_ps(params->avx2_rr1_p5.c5);
31 const __m256 vc4 = _mm256_load_ps(params->avx2_rr1_p5.c4);
32 const __m256 vc3 = _mm256_load_ps(params->avx2_rr1_p5.c3);
33 const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p5.c2);
34 const __m256 vc1 = _mm256_load_ps(params->avx2_rr1_p5.c1);
35 const __m256 vone = _mm256_load_ps(params->avx2_rr1_p5.one);
36 const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx2_rr1_p5.denorm_cutoff);
37
38 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
39 const __m256 vx = _mm256_loadu_ps(x);
40 x += 8;
41
42 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
43
44 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
45 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
46 vn = _mm256_sub_ps(vn, vmagic_bias);
47
48 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
49
50 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
51 vp = _mm256_fmadd_ps(vp, vt, vc3);
52 vp = _mm256_fmadd_ps(vp, vt, vc2);
53 vp = _mm256_fmadd_ps(vp, vt, vc1);
54
55 vt = _mm256_mul_ps(vt, vs);
56 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
57
58 const __m256 vd = _mm256_add_ps(ve, vone);
59 __m256 vr = _mm256_rcp_ps(vd);
60 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
61 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
62 __m256 vf = _mm256_mul_ps(ve, vr);
63
64 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
65 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
66
67 _mm256_storeu_ps(y, vf);
68 y += 8;
69 }
70 if XNN_UNLIKELY(n != 0) {
71 assert(n >= 1 * sizeof(float));
72 assert(n <= 7 * sizeof(float));
73 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2_rr1_p5.mask_table[7] - n));
74
75 const __m256 vx = _mm256_maskload_ps(x, vmask);
76
77 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
78
79 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
80 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
81 vn = _mm256_sub_ps(vn, vmagic_bias);
82
83 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
84
85 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
86 vp = _mm256_fmadd_ps(vp, vt, vc3);
87 vp = _mm256_fmadd_ps(vp, vt, vc2);
88 vp = _mm256_fmadd_ps(vp, vt, vc1);
89
90 vt = _mm256_mul_ps(vt, vs);
91 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
92
93 const __m256 vd = _mm256_add_ps(ve, vone);
94 __m256 vr = _mm256_rcp_ps(vd);
95 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
96 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
97 __m256 vf = _mm256_mul_ps(ve, vr);
98
99 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
100 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
101
102 __m128 vf_lo = _mm256_castps256_ps128(vf);
103 if (n & (4 * sizeof(float))) {
104 _mm_storeu_ps(y, vf_lo);
105 vf_lo = _mm256_extractf128_ps(vf, 1);
106 y += 4;
107 }
108 if (n & (2 * sizeof(float))) {
109 _mm_storel_pi((__m64*) y, vf_lo);
110 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
111 y += 2;
112 }
113 if (n & (1 * sizeof(float))) {
114 _mm_store_ss(y, vf_lo);
115 }
116 }
117 }
118